Artificial intelligence (AI) tools are proliferating in health care at a velocity and scale that are unlike any recent technological or biomedical innovations. ChatGPT launched in 2022 along with other conversational AI chatbots, as did OpenEvidence, a search engine with citations from over 35 million peer-reviewed publications, enabling clinicians to access just-in-time, evidence-based answers to clinical questions. In the ensuing few years, we have ushered in a vital if volatile, wave of evidence generation and mobilization tools that have touched every part of the health care landscape. Rapid adoption has also created opportunities for health services researchers to understand how AI is transforming healthcare in subtle and unsubtle ways. To this end, we invited papers for a thematic issue on the impact of AI on health services research (HSR) from the membership of AcademyHealth, the professional society for HSR. The society's membership represents a wide range of professionals who work in a variety of settings, such as large academic health systems, community-based organizations, public health, and industry. These professionals generate policy-relevant insights about health data, health care structure and infrastructure, and workforce characteristics applied to a range of health topics, conditions, and populations. The research reports and experience reports in this supplemental issue of the LHS Journal illuminate the transformative potential—and transformative realities—of AI. A wide and growing range of AI and machine learning methods, tools, and applications are contributing to how learning health systems generate and apply evidence, and how care is organized and delivered. For example, applications of AI can support predictive modeling for specific at-risk populations as illustrated by Olchanski and colleagues 1. With access to larger and larger EHR-derived data networks 2, the predictive ability and precision of these estimates can improve on prior training datasets for foundation models that were built on data with embedded biases based on demographic characteristics 3. Intentionally deployed algorithms can complement existing manual and computerized processes, as Savitz and colleagues show in screening for community health worker support needs 4, but concerted attention to the trustworthiness of the tools is paramount, which Vald and colleagues demonstrate in their research on nursing handoffs and documentation 5. The unique range of topics in this supplement spans critical arenas including advanced care planning 6, identification of safety-related adverse events 7, and perception of AI in health care from under-represented groups 8 and illustrate the importance of garnering diverse perspectives not only from clinical teams, but also from patients and communities. Validating and evaluating AI tools while they are simultaneously being adopted in practice is inherently challenging. Often, advances in health services and health policy research unfold in response to, or in conjunction with, regulatory and legislative shifts. The federal and state regulatory dialogue around AI is dynamic and unsettled; however, frameworks such as the National Academy of Medicine AI Code of Conduct 9 and the Digital Medicine Society's AI in health care implementation playbook 10 can serve as field reference points. Many health systems are developing local guidelines and guardrails in real time, yet these may not capture performance data on the AI applications when deployed at scale. Given the pace at which AI adoption is occurring, the foresight represented in this compilation is instructive for the field of delivery system science and learning in health and healthcare. Researchers must continue to work with agility and rigor to capture lessons as AI roll-outs roll on. This is truly an example of catching a wave, and the prospect of a wipeout should be foregrounded as researchers study the integration, implementation, and effectiveness of AI in various settings. How we ride this wave is critical to continuously improving care and outcomes. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
Greene et al. (Fri,) studied this question.